concilium-brainstorm

Run a brainstorming campaign on: $ARGUMENTS

Brainstorming is forge's open-book inversion, for a question where the conventional methods are exhausted and nobody knows where the door is. Seats get a LIVE READ-ONLY copy of the system under discussion — its data and/or its source tree, whichever exist; volume is quotaed by boldness tier; every idea carries re-executable evidence; ideas end in human-reviewed dossiers, never an orchestrator shortlist. Read references/brainstorming-mode.md FIRST and follow its evidence tiers — wire nothing marked [open] as a rule. The governing pipeline rules live in references/modes.md: §"IDEAS are never synthesised", §"A research session does NOT implement", §"Every multi-seat round ends in a SYNTHESIS" (findings only).

If $ARGUMENTS is a question (new campaign)

1. Instrument before brief. An isolated guest holds a throwaway READ-ONLY replica of whatever the system has: source tree as a private copy per seat; data as an immutable snapshot each seat can query but not damage. When the data lives in a database, the proven wiring is a rootless Postgres restore with ONE SELECT-only role per seat (default_transaction_read_only = on, statement timeout) and log_statement='all' so the server's own log is the per-seat dig log; for file-shaped data, read-only copies plus shell-history capture play the same roles. Fixture freshness is a validity condition — gate every round on a fixture-vs-live drift check; a stale snapshot has manufactured a confident three-vendor convergent falsehood.
2. Write the brief with, non-negotiably: the boldness quota (4 incremental / 4 aggressive / 4 "laughed out of the room", tier-3s stating why_absurd + what_must_be_true — fewer than 4 tier-3 is a failed run); the evidence contract (the query or command you RAN and what it RETURNED — both are replayed); the denominator rule (every rate names its population); THREE deliverables (ideas, FINDINGS — the fast payoff, resource REQUESTS); the known-intentional-absences list (or seats rediscover them every round); ground-truth boundaries (what in the data is the system's own output and therefore never a correctness signal).
3. Run the seats in parallel where transports are disjoint. Big packets go to the seat as a FILE in its cwd with a pointer prompt (Linux caps one argv string at 128 KiB). Checkpoint-as- you-go is mandatory in every prompt. Verify ARTIFACTS, never exit codes or markers; retry a failed seat once; salvage order is final-answer vs checkpoint, richer wins.
4. Replay every evidence query/command against the same fixture and score. Then write the FINDINGS synthesis (convergent / disagreements / single-seat / retracted — the disagreement section is mandatory and the payoff), verify the load-bearing findings fixture-side, and spawn implementation sessions for anything touching live — the brainstorming session never verifies against live "just this once".